Knowledge graph fraud-related subject analysis and completion method based on large language model
A large language model-based method for analyzing and completing fraud-related subjects in a knowledge graph addresses the challenge of incomplete fraud link information by enhancing entity relationship identification and connection, effectively uncovering complex fraud networks.
Patent Information
- Application Number
- CN202410912488.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-09
- Publication Date
- 2024-12-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing fraud analysis methods fail to explore relationships between fraud subjects in depth, leading to incomplete fraud link information and missing data connections, especially in complex fraud ecosystems.
A method utilizing a large language model to analyze and complete fraud-related subjects in a knowledge graph by cleaning data, extracting entities and relationships, optimizing the graph, and performing link reasoning to fill in missing information.
Enhances the ability to discover and complete fraud-related subjects by constructing a complex knowledge graph that identifies and connects entities with strong representation capabilities, addressing the issue of incomplete fraud link information.
Smart Images

Figure CN119066132A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of electronic digital data processing, and in particular to a method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model. Background Art
[0002] Fraud subjects are key subjects in online fraud activities and play a key role in spreading illegal content on the Internet. Usually, mature fraud subjects have strict organizational structures and collaborative operation models. These fraud resources and fraud teams are intertwined to form a large-scale fraud ecosystem. It is difficult to sort out their relationships through traditional single-dimensional research methods, so it is very suitable to study them through the graph structure of knowledge graphs. Although the academic community has begun to explore the identification of fraud subjects, the in-depth analysis of the internal correlation of fraud resources is still in its initial stage, mainly focusing on the analysis of some single-type fraud subjects, and no correlation analysis and mining has been conducted.
[0003] Among them, the patent [CN117675348A] mainly analyzes fraudulent SMS based on knowledge graph, and uses the similarity calculation method based on the association analysis function of the knowledge graph to predict whether the target SMS port is a new fraudulent SMS port; Patent
[0004]
CN116521892A
CN114124576A
CN113254844A
CN113641827A
CN112035677A
[0005] The above not only fails to explore the relationship between the fraud subjects in depth, but also the fraud methods of the fraud subjects are relatively single, and the connection between different fraud methods of the same gang is not excavated. In the conversion process, there are inevitably missing links related to fraud, which need to be manually completed. It cannot cope with the situation of missing data and complex connections in real life. Summary of the invention
[0006] The present invention aims to solve the problem of incomplete fraud link information and provides a method for analyzing and completing fraud-related subjects based on a knowledge graph based on a large language model. The knowledge graph completed based on a large language model is applied to the fields of fraud-related relationship completion and fraud-related subject discovery. The relationship network between fraud-related entities is explored through similarity, and the model is used to complete entity relationships and analyze fraud subjects.
[0007] The present invention provides a method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model, comprising the following steps:
[0008] S1. Clean the originally collected fraudulent website samples and normal website samples, remove the randomly changed or randomly referenced data information to obtain the data to be analyzed;
[0009] S2. Extract fraud-related entities from the data to be analyzed and sort out their relationships to obtain fraud-related knowledge graph entity-relationships, and then map the knowledge graph to the graph database to obtain pre-processed structured data;
[0010] S3. Optimize the knowledge graph, perform entity fusion and remove redundancy on the preprocessed structured data so that each entity appears only once in the knowledge graph;
[0011] S4, perform link reasoning and link completion on the knowledge graph based on the big prediction model;
[0012] S5. Use the PageRank algorithm to calculate the mutual link weights between each node in the knowledge graph, obtain the relationship weight of each node in the global graph, and then calculate the similarity between the nodes to find all unknown fraud-related nodes that are similar to known fraud-related nodes. A knowledge graph fraud subject analysis and completion method based on a large language model is completed.
[0013] The method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model described in the present invention is, as a preferred embodiment, step S1 includes the following steps:
[0014] S11. Label the fraudulent website samples and normal website samples, use crawlers to crawl the record information and IP addresses of the fraudulent website samples and normal website samples, and crawl the website source code; obtain fraudulent SMS data, normal SMS data, fraudulent app data and normal app data from known or public traffic data and obtain the original data;
[0015] S12, cleaning the website data in the original data, and then cleaning the source code, determining the MD5 value of the website framework structure, so as to remove the data information that changes randomly or is randomly referenced; then dialing the IP address of the website of the original data, confirming the IP location, crawling the website's filing information and obtaining the data to be analyzed;
[0016] S13. Map the data to be analyzed to the graph database through the designed fraud-related knowledge graph and save it in the relational database neo4j.
[0017] The method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model described in the present invention is, as a preferred method, in step S12, a method for cleaning website data is: cleaning the website text content and retaining only the tag information;
[0018] The source code cleaning method is to clean the text information contained in the website source code and only keep all the tags and response header names in the source code.
[0019] The method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model described in the present invention is, as a preferred embodiment, in step S2, the entity-relationship of the fraud-related knowledge graph includes: a head node entity h, a relationship r, and a tail node entity t;
[0020] Relationship r includes: located at, filed at, linked to, belongs to, based on, owned by, points to;
[0021] The corresponding head node entity h and tail node entity t are: IP-IP dialing address, website-registration number, SMS-website, SMS short link-SMS body, website-website source code framework md5 value, IP dialing address including domestic / international, city name;
[0022] When the relationship r is the owner, the corresponding head node entity h and tail node entity t include: website-fraud subject, App-fraud subject, SMS-fraud subject;
[0023] When the relationship r is pointing, the corresponding head node entity h and tail node entity t include: SMS-website, website-ip;
[0024] When the relationship r is similarity, the corresponding head node entity h and tail node entity t include: website-website, subject-subject.
[0025] The method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model described in the present invention is, as a preferred method, in step S2, identifying the fraud-related named entities in the data to be analyzed, and associating and merging the entities with the same name to obtain the fraud-related knowledge graph entity-relationships.
[0026] The method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model described in the present invention is, as a preferred method, to perform fraud-related named entity recognition using a named entity recognition and relationship extraction strategy based on a large language model ChatGLM;
[0027] For text messages, the large language model ChatGLM is required to automatically return the entity name when input. The model automatically extracts the subject name in the text message and the app. The app extracts the app name displayed in the package, and the text message uses the large model questioning method.
[0028] The method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model described in the present invention is, as a preferred method, in step S3, the method for optimizing the knowledge graph is: removing redundancy from the preprocessed structured data (h, r, t), when the same named entity points to different text messages and websites, calculating entity similarity and marking the similarity of the website entity between the website entities, merging and removing duplicates of the related relationships of entities with the same name;
[0029] The method for calculating entity similarity is:
[0030]
[0031] Among them, θ is the entity similarity, a and b are both website entity feature vectors;
[0032] The website entity feature vector includes the concatenated text vector embedding and one-hot vector information;
[0033] The method of converting text vector embedding is as follows: extract the text embedding information of the website entity name based on the Word2vec model and convert it into a vector;
[0034] The method for converting one-hot vector information is as follows: convert the source code, IP and registration number associated with the website entity into one-hot vector information of fixed length 20 to 50 dimensions and accumulate them in a ratio of 8:1:1.
[0035] The method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model described in the present invention is, as a preferred method, in step S4, the method for completing links in the knowledge graph is: using a large language model to perform small sample pre-training, inputting a batch of known samples, known fraud-related domain names and registration numbers, known partial fraud-related subjects and a large number of similar website addresses, and the relationship between the fraud-related subjects and known apps, and predicting partial relationships;
[0036] The training samples of the large language model are a large number of known triple information. The triple information is the known website and entity affiliation. The website and entity affiliation includes: the correspondence between IP and IP address, the affiliation between app and subject, the pointing relationship between SMS and website, and the similarity relationship and value between websites.
[0037] The method for predicting partial relations is: given a triple (h, r, t) and writing a fixed prompt sentence, the task query is to classify the triple (h, r, t) as correct or incorrect. The method for judging the correct classification is: the entity similarity is greater than the threshold; the links are completed in turn according to the judgment results and stored in the graph database.
[0038] The method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model described in the present invention is preferably implemented with a threshold of 0.85.
[0039] The method for analyzing and completing fraud-related entities in a knowledge graph based on a large language model described in the present invention is, as a preferred method, in step S5, a link relationship graph between website entities is constructed and iterative calculation is performed until convergence; in each iteration, the score of each web page is recalculated and updated to the next iteration; finally, the search results are sorted according to the size of the entity score, and the web pages ranked in the top few are output as fraud-related entities;
[0040] The PageRank value of each point is equal to the PageRank value of its entry point divided by the weighted sum of the number of outgoing edges of the entry point. When there are four website entities A, B, C, and D, the PageRank score PR(A) of point A in each round is:
[0041] PR(A)=d*(PR(D) / 2+PR(B) / 1+PR(C) / 2)+(1-d);
[0042] Among them, d is a correction coefficient between 0 and 1, PR(B) is the score of the PageRank value of point B, PR(C) is the score of the PageRank value of point C, and PR(D) is the score of the PageRank value of point D.
[0043] The present invention has the following advantages:
[0044] (1) A fraud-related entity uses a batch of devices to register a large number of fraud-related websites and create corresponding apps, and sends a large amount of false information by impersonating base station services. The information contains short links that directly point to the fraud-related websites or to the download websites of the fraud-related apps. Fraud-related gangs are highly anonymous, and criminals often deliberately hide their identity information, such as location, service provider, etc. In the knowledge graph, anonymity is manifested as the lack of relationships. By constructing a sufficiently complex knowledge graph, the present invention extracts entities and relationships that have strong representation capabilities for the identity information of fraud-related assets as much as possible, and uses knowledge graph learning to complete the missing information.
[0045] (2) Existing solutions start with a certain type of fraud method, or analyze the fraud subject through the case information and financial records of the fraud gang. They lack the mining and analysis of the relationship between different fraud methods, and it is difficult to deal with the problem of incomplete fraud link information in real life. Therefore, this technical solution innovatively applies the knowledge graph based on large language model completion to the field of fraud-related relationship completion and fraud-related subject discovery. It explores the relationship network between fraud-related entities through similarity, and uses the model to complete the entity relationship and analyze the fraud subject. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 This is a flowchart of a method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model;
[0047] Figure 2 It is an entity relationship diagram of a fraud-related knowledge graph for a fraud-related subject analysis and completion method of a knowledge graph based on a large language model;
[0048] Figure 3 This is a schematic diagram of the key node analysis of a knowledge graph based on a large language model for analyzing and completing knowledge graph fraud-related subjects. DETAILED DESCRIPTION
[0049] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.
[0050] Example 1
[0051] like Figures 1 to 3 As shown in the figure, a method for analyzing and completing fraud-related subjects in knowledge graphs based on a large language model is proposed. The method innovatively applies the knowledge graph completed based on a large language model to the fields of fraud-related relationship completion and fraud-related subject discovery. By calculating the similarity of nodes, the model is used to complete entity relationships and analyze fraud subjects.
[0052] Fraud-related knowledge graph ontology design: To better explain fraud-related activities, we can summarize the fraud scenario of the fraud team into the following typical case:
[0053] A fraud-related entity uses a batch of devices to register a large number of fraud-related websites and create corresponding apps, and impersonates base station services to send a large amount of false information. The information contains short links that directly point to the fraud-related websites or to the fraud-related app download websites. Fraud-related gangs are highly anonymous, and criminals often deliberately hide their identity information, such as location, service provider, etc. In the knowledge graph, anonymity is manifested as the lack of relationships. By constructing a sufficiently complex knowledge graph, the knowledge graph representation learning is used to complete the missing information. Therefore, the extracted ontologies and relationships need to have strong representation capabilities for the identity information of the fraud-related assets as much as possible. For example Figure 1 As shown, the ontology required for traceability can be organized as follows:
[0054] (1) Fraudulent entity (2) Fraudulent website (main domain name) (3) Website registration number (4) Website IP (5) Website IP registration location (6) Website source code MD5 value (7) Fraudulent app (8) Fraudulent SMS
[0055]
[0056] The main process of this technical solution is as follows:
[0057] 1. Data acquisition and cleaning: Remove data that does not meet the analysis requirements from the original collected fraud website data.
[0058] 11. Information acquisition module: When some of the fraud samples and normal samples with annotations are known, crawlers are used to crawl the website-related registration information and IP addresses, and crawl the website source code; relevant SMS data and app data are obtained from known or public traffic data.
[0059] 12. Data cleaning:
[0060] First, we clean the website-related information, the original collected fraud website data, and then clean the source code, clean the content data and retain the framework structure code of the web page. Clean the text content of the website and only retain the tag information. The website source code contains a lot of text information, and the text, pictures, and local referenced file names often change randomly. We only need to retain all the tags and response header names in the source code, remove all references and random data information, and determine the md5 value of the website framework structure. Secondly, we dial the IP addresses of all known websites and confirm the IP location; crawl the website's filing information, and finally obtain the website's jump link information, and save the jump and pointing relationship between websites.
[0061] 13. Then the data that meets the analysis requirements is mapped to the graph database through the designed knowledge graph and saved in the relational database neo4j.
[0062] 2. Identification and Relationship Extraction of Fraud-related Entities: Extract entities and sort out relationships, and then map them to a graph database through a designed knowledge graph.
[0063] The specific method for identifying and extracting relationships of fraud-related entities is as follows: Identify named entities, associate entities with the same name, merge their relationships. After entity and link processing, the entity-relationship graph of the fraud-related knowledge graph is as Figure 1 shown. Adopt a named entity recognition and relationship extraction strategy based on the open large language model (ChatGLM). For example, for text messages, when inputting, require the language model to automatically return the entity names. The model can automatically extract the main body names in text messages and apps. For apps, mainly take the app names displayed in the package. For text messages, use the method of asking questions with the large model. Examples of text message extraction questions:
[0064] Question: Please only return the entity names in the following information "Hello, I am your exclusive little sister Tian Tian at Youfa International. The Zhang number is waiting to receive the Hongbao. Add the exclusive to consult for more discounts and click youfaguoji.com".
[0065] Answer: Youfa International.
[0066] Question: Please only return the main body name of the (app name) "You. Fa. Guo. Ji".
[0067] Answer: Youfa International.
[0068] 3. Graph Optimization: First, entity fusion needs to be performed on the preprocessed structured data. Redundancies may occur in fraud-related entities at the initial stage of graph creation. Regardless of the identity attributes of named entities, the graph needs to ensure that they only appear once when constructing the graph. Here, identify the same named entities in the knowledge graph and fuse the links between entities.
[0069] The specific method for graph optimization is:
[0070] After obtaining the structured triples (h, r, t) in the graph, it is still necessary to process the redundant entities and relationships among them.
[0071] First, entity fusion and deduplication need to be performed on the preprocessed structured data. Redundancies may occur in fraud-related entities at the initial stage of graph creation. Regardless of the identity attributes of named entities, the graph needs to ensure that they only appear once when constructing the graph. Here, identify the same named entities in the knowledge graph and fuse the links between entities.
[0072] The specific method is: website entity similarity calculation: merge the related relationships of entities with the same name and remove duplicates. If the same named entity points to different text messages and websites, the entity similarity calculation can be performed: first, based on the website name text, usually the website name is a short text, based on the Word2vec model, extract the embedding information and convert it into a vector, and then convert the associated source code, IP and filing number into a fixed length 20-50 dimension one-hot vector information, accumulate it in a ratio of 8:1:1, and splice it with the corresponding text embedding vector information, calculate the entity similarity based on the cosine similarity, and mark it between the website entities. The cosine similarity formula is as follows:
[0073] Among them, θ is the similarity, and a and b are both website entity feature vectors.
[0074] 4. Graph completion: Perform link reasoning and link completion on the knowledge graph based on the big prediction model.
[0075] The specific method of knowledge graph link completion is: use a large language model and pre-train with small samples, input a batch of known samples, known fraud domain names and registration numbers, known relationships between some fraud entities and a large number of similar URLs, and the relationship between fraud entities and known apps, and predict some relationships.
[0076] (1) The training samples are a large amount of known triple information (a large amount of known website and entity affiliation, IP and IP address correspondence, app and subject affiliation, SMS and website pointing relationship, similarity relationship and value between websites)
[0077] (2) Write a fixed prompt. Given a triple (h, r, t), write a fixed prompt sentence. The task is to classify it as correct or incorrect. For example:
[0078] (Website 1 - Similarity - Website 2), Similarity 0.98, (Website 1 - Owner is - Fraudulent Subject A) Training input, Website 2 owner is Fraudulent Subject A, is this correct? Answer: Yes.
[0079] (SMS-points to-Website 1), (Website 1-owner is-fraud subject a), question such as SMS owner is fraud subject a, is it correct? Answer: Yes, complete the link according to the result, store it in the graph database, and so on.
[0080] 5. Analysis of key fraud-related entities: Use the PageRank algorithm to calculate the link weights between each node in the knowledge graph, obtain the relationship weight of each node in the global graph, and then calculate the similarity between the nodes to find all unknown fraud-related nodes that are similar to known fraud-related nodes.
[0081] The specific method of analyzing the key nodes of the knowledge graph is: use the PageRank algorithm to calculate the connection weights between each subject in the knowledge graph, and obtain the relationship weight of each subject in the global graph; the specific steps to implement the PageRank algorithm include: first construct a link relationship graph between website subjects, and then iterate the graph until convergence. In each iteration, the score of each web page will be recalculated and updated to the next iteration. Finally, sort the search results according to the size of the entity score and output the top-ranked web pages. There are 4 subjects as follows, A, B, C, D:
[0082] Taking point A as an example, the calculation method of its PageRank value in each round is as follows: PR(A)=d*(PR(D) / 2+PR(B) / 1+PR(C) / 2)+(1-d);
[0083] The PageRank value of each point is equal to the PageRank value of its entry point divided by the weighted sum of the number of outgoing edges of the entry point, where d is a correction coefficient between 0 and 1.
[0084] The specific usage method of this embodiment is as follows:
[0085] The entity named "Treasury Bonds" has hundreds of websites with different domain names. The content frameworks of the websites are similar or similar, and there is no registration information. There are also corresponding apps. Using a knowledge graph fraud subject analysis and completion method based on a large language model, the entity of "Treasury Bonds" is calculated as a key node and focused on and analyzed;
[0086] Once a new unknown website, SMS or app comes in, the system will use the questions and answers predefined by the big model to inquire whether it points to these key fraud-related entities. If it points to entity data, the analysis system will focus on it while completing the link to the problematic entity, indicating that the fraud-related team has registered a new domain name or developed a new promotion method.
[0087] If there are similar entities connected to different websites, apps, and text messages, they can also be aligned and merged based on questions and answers. For example, if the large model is asked how relevant the name of the fraud-related entity is, the answer is a similarity between 0-1. Based on the previous training data, the model obtains a similarity of 0.87, and the two entities can be merged, and the website relationship links can be aligned to complete them, enriching and optimizing the existing network structure.
[0088] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
Claims
1. A method for analyzing and completing fraud-related subjects in knowledge graphs based on a large language model. Features: The following steps are involved: S1. Clean the originally collected fraudulent website samples and normal website samples, remove the randomly changed or randomly referenced data information to obtain the data to be analyzed; S2. Extracting and sorting out the fraud-related entities in the data to be analyzed to obtain the fraud-related knowledge graph entity-relationship, and then mapping the knowledge graph to the graph database to obtain the preprocessed structured data; S3, optimizing the knowledge graph, performing entity fusion and removing redundancy on the preprocessed structured data, so that each entity appears only once in the knowledge graph; S4, perform link reasoning and link completion on the knowledge graph based on the big prediction model; S5. Use the PageRank algorithm to calculate the mutual link weights between each node in the knowledge graph, obtain the relationship weight of each node in the global graph, and then calculate the similarity between the nodes to find all unknown fraud-related nodes that are similar to known fraud-related nodes. A knowledge graph fraud subject analysis and completion method based on a large language model is completed.
2. According to the method of analyzing and completing fraud-related subjects in a knowledge graph based on a large language model according to claim 1, Features: Step S1 includes the following steps: S11. Label the fraudulent website samples and normal website samples, use a crawler to crawl the record information and IP addresses of the fraudulent website samples and the normal website samples, and crawl the website source code; obtain fraudulent SMS data, normal SMS data, fraudulent app data, and normal app data from known or public traffic data and obtain the original data; S12, cleaning the website data in the original data, and then cleaning the source code, determining the MD5 value of the website framework structure, so as to remove the data information that changes randomly or is randomly referenced; then dialing the IP address of the website of the original data, confirming the IP location, crawling the website's filing information and obtaining the data to be analyzed; S13. Map the data to be analyzed to a graph database through a designed fraud-related knowledge graph and save it in a relational database neo4j.
3. According to claim 2, a method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model, Features: In step S12, the website data cleaning method is: cleaning the website text content and only retaining the tag information; The source code is cleaned in the following manner: the text information contained in the website source code is cleaned and only all tags and response header names in the source code are retained.
4. According to the method of analyzing and completing fraud-related subjects in a knowledge graph based on a large language model according to claim 1, Features: In step S2, the entity-relationship of the fraud-related knowledge graph includes: a head node entity h, a relationship r, and a tail node entity t; The relationship r includes: located at, filed at, linked to, belongs to, based on, owned by, and points to; The corresponding head node entity h and tail node entity t are: IP-IP dialing address, website-registration number, SMS-website, SMS short link-SMS body, website-website source code framework md5 value, the IP dialing address includes domestic / international, city name; When the relationship r is the owner, the corresponding head node entity h and the tail node entity t include: website-fraud subject, App-fraud subject, SMS-fraud subject; When the relationship r is pointing, the corresponding head node entity h and tail node entity t include: SMS-website, website-ip; When the relationship r is similarity, the corresponding head node entity h and the tail node entity t include: website-website, subject-subject.
5. According to claim 1, a method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model, Features: In step S2, the fraud-related named entities in the data to be analyzed are identified, and entities with the same name are associated and their relationships are merged to obtain the fraud-related knowledge graph entity-relationship.
6. According to claim 5, a method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model, Features: Use named entity recognition and relation extraction strategies based on the large language model ChatGLM to identify fraudulent named entities; For text messages, the large language model ChatGLM is required to automatically return the entity name when input. The model automatically extracts the subject name in the text message and the app. The app extracts the app name displayed in the package, and the text message uses the large model questioning method.
7. According to claim 1, a method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model, Features: In step S3, the knowledge graph optimization method is as follows: redundancy is removed from the preprocessed structured data (h, r, t); when the same named entity points to different text messages and websites, entity similarity is calculated and the similarity of the website entity is marked between the website entities; the related relationships of the entities with the same name are merged and deduplicated; The method for calculating the entity similarity is: Among them, θ is the entity similarity, a and b are both website entity feature vectors; The website entity feature vector includes concatenated text vector embedding and one-hot vector information; The text vector embedding conversion method is: extracting text embedding information of website entity names based on the Word2vec model and converting it into a vector; The method for converting the one-hot vector information is as follows: converting the source code, IP and registration number associated with the website entity into one-hot vector information of fixed length 20 to 50 dimensions and accumulating them in a ratio of 8:1:
1.
8. According to claim 1, a method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model, Features: In step S4, the method for completing the knowledge graph link is: using a large language model to perform small sample pre-training, inputting a batch of known samples, known fraud-related domain names and registration numbers, known partial fraud-related entities and a large number of similar URL relationships, and the relationship between the fraud-related entities and known apps, and predicting partial relationships; The training samples of the large language model are a large number of known triple information, which are the known website and entity affiliation relationships, including: the correspondence between IP and IP address, the affiliation between app and subject, the pointing relationship between SMS and website, and the similarity relationship and value between websites; The method for predicting partial relations is: given a triple (h, r, t) and writing a fixed prompt sentence, the task query is to classify the triple (h, r, t) as correct or incorrect. The method for judging the correct classification is: the entity similarity is greater than the threshold; the links are completed in turn according to the judgment results and stored in the graph database.
9. According to claim 8, a method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model, Features: The threshold is 0.
85.
10. According to claim 1, a method for analyzing and completing fraud-related subjects in a knowledge graph based on a large language model, Features: In step S5, a link relationship graph between website entities is constructed and iterative calculation is performed until convergence; in each iteration, the score of each web page is recalculated and updated in the next iteration; finally, the search results are sorted according to the size of the entity score, and the web pages ranked in the top few are output as the fraud-related entities; The PageRank value of each point is equal to the PageRank value of its entry point divided by the weighted sum of the number of outgoing edges of the entry point. When there are four website entities A, B, C, and D, the score PR(A) of the PageRank value of point A in each round is: PR(A)=d*(PR(D) / 2+PR(B) / 1+PR(C) / 2)+(1-d); Among them, d is a correction coefficient between 0 and 1, PR(B) is the score of the PageRank value of point B, PR(C) is the score of the PageRank value of point C, and PR(D) is the score of the PageRank value of point D.
Citation Information
Cited By
Metacognition and multi-angle prompt-based large language model knowledge graph completion method
CN119398155A
Material data cleaning method and system fusing large language model and knowledge graph
CN120744327A
High-risk call identification method and device, computer equipment and storage medium
CN120935298A
Inference method, system and equipment of wide constraint large language model and medium
CN121146067A
Abnormal entity information extraction method and system based on large language model
CN122334258A
